Skip to content
NTPU AI Learn

AI 104 · Module 2

Task Decomposition & Execution Steps

How should a complex job be broken into steps an agent can execute?

Start Section 1
  • Duration30 min
  • LevelCore
  • PrerequisiteAI 103 recommended, plus Module 1 of this course
0 / 4 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Explain in order how Goal, Major Tasks, Subtasks, Inputs / Outputs and Dependencies relate
  • Answer six design questions for each step: objective, input, output, precondition, who executes it and how completion is verified
  • Distinguish a Human Task, an AI Task and a Tool Step, and mark the steps that need an external tool
  • Break your own task into 5 to 10 execution steps and add Goal, Inputs, Tasks and Outputs to the Agentic AI Workflow Canvas

Decomposition model

8 min

From Goal to Dependencies

"Summarize course feedback" sounds like one thing, but for an agent to execute it, it must be broken into steps that each have a clear input and output. Task Decomposition means cutting a complex job, layer by layer, into small steps that can be executed and checked.

Task Decomposition model
  1. Goal: what should this work achieve?
  2. Major Tasks: which main pieces of work achieve the goal?
  3. Subtasks: which small steps does each major task break into?
  4. Inputs / Outputs: what does each step need and what does it produce?
  5. Dependencies: which step must wait for another step to finish?

Text equivalent: the five layers run from top to bottom. First set the Goal, then cut it into Major Tasks and each Major Task into Subtasks; next write the Input and Output of every step; finally mark which steps depend on which.

Example: weekly course-feedback report (simulated). Goal: produce a weekly report a manager can review
StepMajor TaskInputOutputDependency
S1Collect: get this week's feedbackFeedback data file (simulated)A cleaned feedback listNone
S2Organize: classifyFeedback list, classification criteriaA category for each commentNeeds the output of S1
S3Organize: find issues needing follow-upClassification resultFollow-up issue listNeeds the output of S2
S4Write: produce the draft reportClassification result, follow-up list, report formatDraft reportNeeds the output of S2 and S3
S5Review: check and confirmDraft report, original feedbackReviewed reportNeeds S4; owned by a person
Check yourself

In the example, S4 "produce the draft report" cannot start until S2 and S3 are finished. What is this relationship called?

In the example, S4 "produce the draft report" cannot start until S2 and S3 are finished. What is this relationship called?

Key takeaway: Start with the Goal, break it into Major Tasks and Subtasks, write the Input and Output of every step, then mark the Dependencies.

Designing each step

8 min

Six questions for every step

Listing steps is only the start. For a step to be executable and checkable, the designer answers six questions about it. The places where you cannot answer are usually where the flow will fail later.

  1. 1. Objective
    What should this step achieve? Ideally a step has one main action. That is the task boundary.
  2. 2. Input
    What data or previous output is needed before starting?
  3. 3. Output
    What will it produce when done? The next step must be able to use it.
  4. 4. Precondition
    What must be true before this step begins? For example: the feedback data is loaded. There should also be a postcondition: what should be true afterward?
  5. 5. Who executes it?
    Is it a Human Task, an AI Task or a Tool Step?
  6. 6. How is completion verified?
    How do you know the step was done correctly? For example: a rule check, comparison with the source, or a person's review.
Three kinds of executor
TypeSuits steps thatExample (simulated scenario)
Human Taskneed accountable judgment, approval, or involve sensitive or high-impact decisionsReviewing the report, deciding whether to send it to the manager
AI Taskare organizing, classifying, summarizing or drafting work that can be reviewedClassifying each comment, producing the draft report
Tool Stepneed an external system, data source or actionGetting the feedback data from the course system, sending an email

Info

In this course, when a step needs an external tool, just mark it as a Tool Step; you do not need to design how the tool is connected. How to connect tools and trusted data safely is the subject of AI 201.
Check yourself

A step needs to look up this week's enrollment count in the school's student data system. How should it be marked?

A step needs to look up this week's enrollment count in the school's student data system. How should it be marked?

Key takeaway: For every step you should be able to state the objective, input, output, precondition, executor and how it is verified.

Practice and apply

8 min

Practice: Task Decomposition

Take the task you chose in Module 1, or pick one of the scenarios below, and break it into 5 to 10 execution steps. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise; describe everything with simulated data.

  • Preparing a course
  • Reviewing an ESG report
  • Analyzing a customer request
  • Preparing an investment brief
  • Reviewing a policy document
  • Processing an administrative request
One row per step (field template)
StepObjectiveInputOutputPreconditionExecutor (Human / AI / Tool)How verified
S1(one main action)
S2
...
  1. Check 1: number of steps
    5 to 10 in total. Fewer means still too coarse; more means cut too finely.
  2. Check 2: one action per step
    A step that "analyzes and writes and sends" should be split.
  3. Check 3: Outputs fit Inputs
    Each step's input must come from the user, a previous output, or a clearly marked data source.
  4. Check 4: where is the person?
    Mark at least one step that a person reviews or approves.
Check yourself

A step is written as "analyze feedback, write the report and send it to the manager". What most needs improving?

A step is written as "analyze feedback, write the report and send it to the manager". What most needs improving?

Key takeaway: A good decomposition has 5 to 10 steps, one action per step, outputs that fit inputs, and people where they belong.

6 min

Apply: put Goal, Inputs, Tasks and Outputs on the Canvas

Done

The result of this module is the first half of your Agentic AI Workflow Canvas. Organize your task decomposition into it.
What to add to the Canvas from this module
Canvas sectionWhat to writeExample (simulated)
GoalOne sentence on what the work should achieveProduce a weekly feedback report a manager can review
Trigger / InputWhat starts the workflow? Which inputs are needed?Every Friday; this week's feedback data file (simulated)
Task DecompositionList Major Tasks and Subtasks (5 to 10 steps)Collect, classify, find follow-ups, draft, review
Execution OutputWhat does each step produce?Feedback list, categories, follow-up list, draft report, reviewed report

Module 1: Workflow or Agent?

  • Judge who decides the next step
  • Choose the task to use throughout the course

Module 2: How to decompose a task?

  • From Goal down to Subtasks
  • Write Input, Output and executor for every step

Human-AI-Human: a person defines the Goal and the step boundaries, the AI executes the steps assigned to it, and a person reviews and stays responsible for the result. The next module asks: must these steps always run in a fixed order?

Key takeaway: Write the Goal, Inputs, Tasks and Outputs clearly first; the later patterns and decisions depend on them.

Save as PDF

What next?